Vol 13 No 1 - May 2014

ISSN: 1694-2507 (Print)
ISSN: 1694-2108 (Online)
International Journal of Computer Science
and Business Informatics
(IJCSBI.ORG)
VOL 13, NO 1
MAY 2014

Table of Contents VOL 13, NO 1 MAY 2014
A Novel Facial Recognition Method using Discrete Wavelet Transform Multiresolution Pyramid..........1
G. Preethi
Enhancing Energy Efficiency in WSN using Energy Potential and Energy Balancing Concepts ................. 9
Sheetalrani R. Kawale
DNS: Dynamic Network Selection Scheme for Vertical Handover in Heterogeneous Wireless Networks
.................................................................................................................................................................... 19
M. Deva Priya, D. Prithviraj and Dr. M. L Valarmathi
Implementation of Image based Flower Classification System................................................................ 35
Tanvi Kulkarni and Nilesh. J. Uke
A Survey on Knowledge Analytics of Text from Social Media.................................................................. 45
Dr. J. Akilandeswari and K. Rajalakshm
Progression of String Matching Practices in Web Mining – A Survey ..................................................... 62
Kaladevi A. C. and Nivetha S. M.
Virtualizing the Inter Communication of Clouds ...............................................................................72
Subho Roy Chowdhury, Sambit Kumar Patel, Ankita Vinod Mandekar and G. Usha Devi
Tracing the Adversaries using Packet Marking and Packet Logging ....................................................... 86
A. Santhosh and Dr. J. Senthil Kumar
An Improved Energy Efficient Clustering Algorithm for Non Availability of Spectrum in Cognitive Radio
Users ....................................................................................................................................................... 101
IJCSBI.ORG

V. Shunmuga Sundaram and Dr S. J. K Jagadeesh Kumar

International Journal of Computer Science and Business Informatics
IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 1
A Novel Facial Recognition Method
using Discrete Wavelet Transform
Multiresolution Pyramid
G. Preethi
PG Scholar, Department of CSE,
Chendhuran College of Engineering & Technology,
Pudukkottai – 622507, India
ABSTRACT
Necessity for the facial recognition methods is increasing now-a-days as large number of
applications need it. While implementing the facial recognition methods the cost of data
storage and data transmission plays a vital role. Hence facial recognition methods require
image compression techniques to full fill the requirements. Our paper is based on the
discrete wavelet transform multiresolution pyramid. Various resolutions of the original
image with different image qualities can be had without employing any image compression
techniques. Principal Component Analysis is used to measure the facial recognition
performance using various resolutions of the image. Facial images for testing are selected
from standard FERET database. Experimental results show that the low resolution facial
images also performs equal to the higher resolution images. So instead of using all the
available wavelet coefficients, the minimum number of coefficients representing the lower
resolution can be used and there is no need of image compression.
Keywords
Principal component analysis, discrete cosine transform, discrete wavelet transform,
support vector machine words.
1. INTRODUCTION
Facial recognition methods are used to identify or verify an individual using
the facial images already enrolled in a database. The general categories of
facial recognition are holistic, feature-based, template-based and part-based
methods. Among them holistic method requires the whole face region as
input and utilizes its statistical moments. The basic and commonly used
holistic methods are based on Principal Component Analysis (PCA) [1].
Facial recognition methods are used in large number of applications like e-
visa, e-passport, entry control in organizations, criminal identification,
forensic science, smart phones and laptops for authentication etc. The
number of facial images to be stored increases the problems like data
storage and the cost of transmitting images. As a solution to reduce both
data storage and cost of transmission, image compression algorithms are
utilized.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 2
Efficient image compression can be achieved using transform based
methods than the pixel based methods. Transform coding transforms the
given image from spatial domain to transform domain where efficient
compression can be carried out. Since the transformation is a linear process,
there will not be any loss of information and the number of coefficients
equals the number of pixels. As most of the image’s energy is concentrated
within a few large magnitude coefficients, the remaining very small
magnitude coefficients can be coarsely quantized or even ignored while
encoding. This will not affect the quality of the reconstructed image more.
The available mathematical transforms are Karhunen-Loeve (KLT),
Discrete Fourier Transform (DFT), Discrete Cosine Transform (DCT) and
Discrete Wavelet Transform (DWT) [2]. Among them DCT is utilized in
large applications like JPEG and MPEG. Now DWT is replacing the DCT
by its superior quality and various decoding options. Transforms which
operates on the whole image instead of image blocks can avoid blocking
artifacts at low compression rates. DWT decomposes the source signal into
non-overlapping and contiguous frequency ranges called sub bands. The
source sequence is fed to a bank of band pass filters which are contiguous
and cover the full frequency range. This set of output signals are the sub
band signals and can be recombined without degradation to produce the
original signal [3] [4]. Fig.1 shows how a signal is separated into sub bands
using band pass filters.
Figure 1. Sub band decomposition of a signal
When transforming a two dimensional digital image using the band pass
(low pass and high pass) filters, it requires the first transform along
horizontal axis and the second one along vertical axis to decompose the
image into sub bands. The resulting four sub bands are named as LL, LH,
HL and HH of a one level decomposition. LL, LH, HL and HH represents
h1
h0
2
2
h1
h0
2
2
x(n)

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 3
lowest frequencies, vertical high frequencies (horizontal edges), horizontal
high frequencies (vertical edges) and high frequencies in both directions (the
comers) respectively. Fig.2 shows various sub bands separated by a three
level dyadic DWT [5].
Figure 2. Sub bands separated by a three level dyadic DWT.
The multiresolution property [6] of DWT enables the user to have variable
resolutions of the transformed image. While reconstructing the image, for a
3 level transformation, four resolutions (0 to 3) are possible. The LL3 sub
band can reconstruct 0th resolution, LL3, HL3, LH3 and HH3 sub-bands
can reconstruct 1st resolution, LL3, HL3, LH3, HH3, HL2, LH2 and HH2
sub-bands can reconstruct 2nd resolution and LL3, HL3, LH3, HH3, HL2,
LH2, HH2, LL1, HL1, LH1 and HH1 sub-bands can reconstruct the third
resolution.
When an image of dimension 128 x 128 pixels is transformed by DWT for 3
levels, the LH1, HL1, LL1 and HH1 will have a dimension of 64 x 64
pixels. LH2, HL2, LL2 and HH2 are of 32 x 32 pixels and LH3, HL3, LL3
and HH3 will have a dimension of 16 x 16 pixels. Hence the resolution 0
requires 256 (16 x 16) wavelet coefficients, 1 requires 1024 (32 x 32)
wavelet coefficients, 2 needs 4096 (64 x 64) wavelet coefficients and 3
requires the whole 16384 (128 x 128) wavelet coefficients. With this
multiresolution feature of the DWT, we propose a novel facial recognition
method where the available resolutions of the facial image are used instead
of the whole image.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 4
2. MATERIALS AND METHODS
We briefly explain about the FERET database, PCA and the performance
measure Recognition Rate here.
2.1 Database
FERET database is a standard database for testing facial recognition
algorithms. This database is collected by Defense Advance Research
Projects Agency (DARPA) and the National Institute of Standards and
Technology (NIST) of United States of America (USA) from 1993 to 1997
[7]. The total collection counts to 14051 grayscale facial images. Images
are categorized into various groups depending upon the nature as Fa, Fb, Fc,
Dup I and Dup II with 1196, 1195, 194, 722 and 234 images respectively.
Moon and Philips [8] have analysed the computation and performance
aspects of PCA based face recognition using Feret database.
2.2 Image Types
There are three types of images: Gallery images are the collection of facial
images from known individuals which forms the search dataset. Probe
images are the collection facial images of unknown persons to be identified
or verified by matching the gallery images. Training images are the random
collection facial images from all the available categories. These training
images are used to train the PCA algorithm for facial recognition.
2.3 Principal Component Analysis
An applied linear algebra tool used for dimensionality reduction of the given
data set. It decorrelates the second-order statistics of the data. A 2-D facial
image is converted into a single dimensional vector by joining all the rows
one after another having r (row) x c (columns) elements. For M training
images, there will be M single dimensional vectors. A mean centered image
is calculated by subtracting the mean image from each vector. Based on the
covariance matrix of the mean centered image, Eigen vectors are computed.
The basis vectors which represent the maximum variance direction from the
original image are selected as feature vectors. These feature vectors are
named as Eigen faces or face space. It is not necessary that the number of
feature vectors should be equal to the number of training images. Every
image in the gallery image set is projected into the face space and the
weights are stored in the memory. The face to be probed is also projected
into the face space. The distance between the projected probe image
weights and every projected gallery image weight is computed. The gallery
image having the shortest distance will be treated as the recognized face.
Many PCA based face recognition methods are available. Hybrid versions
of PCA and other methods like Gabor wavelets [9], Support Vector
Machine (SVM) Classifiers [10], etc. are used for face recognition.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 5
2.4 Distance Measure
The distance measures are used to compare the similarity between the probe
and gallery images. The distance measure used in our work is L1. Let x
and y are two vectors of size n and d is the distance between the vectors x
and y. L1 distance or City-Block or Manhattan distance is defined as the
sum of the absolute differences between these two vectors x and y. L1
distance is given in the following equation:
2.5 Performance Measure - Recognition Rate (RR)
We adopted the performance measure from Delac et. al. [12]. The
recognition rate is defined as the ratio between the number of probe images
recognized correctly and the total number of probe images used for
recognition. Both the gallery and probe images are projected in the face
space and the individual similarity score of the probe images are calculated.
Distance measure is used to find out the gallery image having higher
similarity with the probe image. If the identified gallery image is exactly
equal to the probe image then it is declared that it is correctly identified. For
example out of 1000 probe images if 786 are correctly identified than the
RR is 786/1000 = 78.6%.
3. PROPOSED METHOD
Facial image sets of Fa, Fb, Fc, Dup I and Dup II from FERET database are
normalized as per the ISO/IEC 19794-5 standard for facial image data using
the algorithm of Somasundaram and Palaniappan [12]. From the resultant
images of the normalization method, the facial features region (area
covering eyes, nose, mouth) is segmented to the dimension of 128 x 128
pixels. Few of the test images are shown in Fig.3.
Figure 3. Few segmented test images from FERET database
Every segmented facial image is de-noised using median filter and the
intensity values are equalized using histogram equalization. These images
are transformed using DWT with Cohen-Daubechies-Feauveau 9/7
(CDF9/7) filter for 3 levels. The wavelet coefficients of LL3 (16 x 16) are
used for the reconstruction of resolution 0. Wavelet coefficients of LH3,

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 6
HL3, LL3 and HH3 (32 x 32) are used for the reconstruction of resolution 1.
All the wavelet coefficients except LH1, HL1 and HH1 (64 x 64) are used to
reconstruct resolution 2. Whole wavelet coefficients representing all the
levels (128 x 128) are used to reconstruct resolution 3. Fig.4 shows the
various resolutions available.
Resolution 0 Resolution 1 Resolution 2 Resolution 3
Figure 4. Various resolutions available for a 3 level DWT decomposition
The FERET image set Fa is used as gallery image set. Sets Fb, Fc, Dup I
and Dup II are used as probe image sets. A training set of 501 images from
FERET data set obtained from the CSU Face Identification Evaluation
System of Colorado State University is used in our experiment. Among
these training images 80% are from gallery images and 20% from Dup I
images. While performing PCA on the training set, it generates 500 Eigen
vectors. Among these 500 Eigen vectors only the top 200 Eigen vectors
(40% of the total Eigen vectors) are selected as basis vectors. These basis
vectors are used with PCA algorithm to generate the PCA face space
(WPCA).
We performed two types of experiments where in the first experiment the
training and gallery images are of resolution 3 and only the probe images
are varied from resolution 3 to resolution 0. For the second experiment all
the gallery and probe images are varied from resolution 3 to 0. These two
experiments are carried over for every individual probe sets Fb, Fc, Dup I
and Dup II. Initially the face spaces are generated using PCA using training
images for every resolution. While carrying out the experiments the gallery
and probe images are projected to the respective face space as per the
requirement. The L1 distance measure is used to find the similarity scores
of the gallery images.
4. RESULTS AND DISCUSSION
The FERET facial images are transformed using DWT using Matlab
(Version 7) software. The PCA face space generation, projection of gallery,
probe image and similarity score computation are also carried out using
Matlab programs. For every experiment the recognition rates are individual
calculated for every probe image using all the resolution levels.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 7
4.1 Experiment 1
The recognition rates of the probe image sets Fb, Fc, Dup I and Dup II for
the resolution levels 3, 2, 1 and 0 with the gallery and training images of
resolution 3 are given in Table 1.
Table 1. Recognition rate for resolution 3 training and gallery images
Image Type Recognition Rate (%)
Res-3 Res-2 Res-1 Res-0
Fb 86.78 86.78 86.61 81.92
Fc 38.66 37.63 32.47 25.77
Dup I 41.83 41.69 40.58 35.73
Dup II 19.66 19.23 18.80 14.96
For Fb image set the resolutions 3,2 and 1, the RR is more or less equal and
the resolution 0 decreases much. For all the resolution levels 3 to 0, the RR
drops significantly in Fc image sets. In the image sets Dup I and Dup II also
the RR resembles the image set Fc. As an overall observation the RR drops
significantly as the resolution decreases.
4.2 Experiment 2
The recognition rates of the probe image sets Fb, Fc, Dup I and Dup II for
the resolution levels 3, 2, 1 and 0 with the gallery and training images of the
same resolution level are given in Table 2.
Table 2. Recognition rate for all the resolutions
Image Type Recognition Rate (%)
Res-3 Res-2 Res-1 Res-0
Fb 86.78 88.03 88.77 88.87
Fc 38.66 41.75 41.24 42.27
Dup I 41.83 42.11 41.13 40.44
Dup II 19.66 20.09 19.52 18.80
When the training, gallery and probe image sets belong to the same
resolution give better results than the first experiment. For Fb image set the
RR increases for resolutions 3, 2, 1 and 0 steadily. The RR of resolutions 2
to 0 differ by a minimum of 1.25% from the resolution 3. The RR of Fc
shows a good difference between the resolution 3 and others. Even the
resolution 3 differs by 3.5% with resolution 0. For the image sets Dup I and
Dup II the RR increases for resolution 2 from 3, but decreases for resolution
1 and 0 than the resolution 3.
Based on the results of the above two experiments, it is evident that the
facial recognition rates of the lower resolution also equals the higher
resolution. So instead of using the overall wavelet coefficients a minimum

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 8
number of coefficients which can give higher recognition rate can be used
without using any image compression.
5. CONCLUSIONS
Our proposed method presents a facial recognition algorithm based on the
resolution scalability of DWT using PCA. The lower resolution images
require very low bit rate when compared to higher resolution images. But
the lower resolution images give recognition rate more or less equal to the
higher resolution images. This can save the cost of transmission time and
data storage. Our method can fulfill the requirements of a basic facial
recognition with low resolution images.
REFERENCES
[1] Turk, M.A., and Pentland, A.P. Face Recognition using Eigenfaces, IEEE Conference
on Computer Vision and Pattern Recognition, (1991), 586-591.
[2] Salamon, D. Data Compression – The Complete Reference, Second Edition, Springer-
Verlag., 2000.
[3] Robi Polikar, The Wavelet Tutorial, http://users.rowan.edu/~polikar/WAVELETS
[4] Wavelet Theory, Department of Cybernetics,http:cyber.felk.cvut.cz.
[5] William, A., Pearlman, and Amir Said, Digital Signal Compression – Principles and
Practice, Cambridge University Press, 2011.
[6] Mallat and Stephane, G. A Theory of Multiresolution Signal Decomposition: The
Wavelet Representation, IEEE Transactions on Pattern Analysis and Machine
Intelligence, 11, 7(1989), 674-693.
[7] Grayscale FERET Database. http://www.itl.nist.gov/iad/humanid/feret/
[8] Moon, H., and Phillips, P.J. Computational and Performance Aspects of PCA-based
Face Recognition Algorithms, Perception, 30 (2001), 303-321.
[9] Cho, H., Roberts, R., Jung, B., Choi, O., and Moon, S. An Efficient Hybrid Face
Recognition Algorithm Using PCA and GABOR Wavelets. International Journal of
Advanced Robotic Systems, 11, 59 (2014), 1-8.
[10]Xu, W., and Lee, E. J. Face Recognition Using Wavelets Transform and 2D PCA by
SVM Classifier, International Journal of Multimedia and Ubiquitous Engineering, 9, 3
(2014), 281-290
[11]Delac, K., Grgic, M., and Grgic, S. Face recognition in JPEG and JPEG2000
Compressed Domain, Image and Vision Computing, 27 (2009), 1108-1120.
[12]Somasundram, K., and Palaniappan, N. Personal ID Image Normalization using
ISO/IEC 19794-5 Standards for Facial Recognition Improvement, Communications in
Computer and Information Science Series, Springer Verlag, 283 (2012), 429-438.
This paper may be cited as:
Preethi, G. 2014. A Novel Facial Recognition Method using Discrete
Wavelet Transform Multiresolution Pyramid. International Journal of
Computer Science and Business Informatics, Vol. 13, No. 1, pp. 1-8.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 9
Enhancing Energy Efficiency in
WSN using Energy Potential and
Energy Balancing Concepts
Sheetalrani R. Kawale
Assistant Professor, Department of Computer Science
Karnataka State Women’s University, Bijapur
ABSTRACT
There are much different energy aware routing protocols proposed in the literature, most of
them focus only on energy efficiency by finding the optimal path to minimize energy
consumption. These protocols should not only aim for energy efficiency but also for energy
balance consumption. In this work, energy balanced data gathering routing algorithm is
developed using the concepts of potential in classical physics [16]. Our scheme called
energy balanced routing protocol, forwards data packets toward the sink through dense
energy areas so as to protect the nodes with relatively low residual energy. This is to
construct three independent virtual potential fields in terms of depth, energy density and
residual energy. The depth field is used to establish a basic routing paradigm which helps in
moving the packets towards the sink. The energy density field ensures that packets are
always forwarded along the high energy areas. Finally, the residual energy field aims to
protect the low energy nodes. An energy-efficient routing protocol, tries to extend the
network lifetime through minimizing the energy consumption whereas energy balanced
with efficiency routing protocol intends to prolong the network lifetime through uniform
energy consumption with efficiently.
Keywords
Sensor networks, energy efficient routing, potential fields, low energy nodes.
1. INTRODUCTION
Recent development in wireless technology has enabled the development of
low power, multifunctional sensor nodes that are in small size and
communicate in small distances. This tiny sensor node, which consists of
sensing, data processing and communicating components, leverage the idea
of sensor networks. A sensor network is composed of a large number of
sensor nodes that are densely deployed either inside the phenomenon or
very close to it. The positions of these sensor nodes can be easily engineered
to be either fixed to a particular location or have certain amount of mobility
in a predefined area. [24][25]

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 10
2. BACKGROUND STUDY
The sensing or monitoring of for example temperature, humidity etc.,
constitutes one of the two main tasks of each sensor. The other main task is
packet forwarding using the equipped wireless technology. Whichever way
data is transmitted the network must provide a way of transporting
information from different sensors to wherever this information is needed.
Sensor networks could be deployed in a wide variety of application domains
such as military intelligence, commercial inventory tracking and agricultural
monitoring [22][23][24].
Each node stores the identity of one or more nodes through which it heard
an announcement that another group exists. That node may have itself heard
the information second-hand, so every node within a group will end up with
a next-hop path to every other group, as in distance-vector. Topology
discovery proceeds in this manner until all network nodes are members of a
single group. By the end of topology discovery, each node learns every
other node’s virtual address, public key, and certificate, since every group
members knows the identities of all other group members and the network
converges to a single group.
3. EXISTING SYSTEM
The existing system focus on energy efficient routing whose target is to find
an optimal path to minimize energy consumption on local nodes or in the
whole WSN [17][18][19]. The energy aware routing maintains multiple
paths and properly chooses one for each packet delivery to improve network
survivability. It may be quite costly since indeed to exchange routing
information very frequently and may result in energy burden and traffic
overload for the nodes.
4. PROBLEM IDENTIFICATION
Energy is an important resource for battery-powered wireless sensor
networks (WSN) that makes energy-efficient protocol design a key
challenging problem. The three main reasons that can cause an imbalance in
energy distribution:
 Topology: The topology of the initial deployment limits the number
of paths along which the data packets can flow. For example, if there
is only a single path to the sink, nodes along this path would deplete
their energy rather quickly. In this extreme case, there are no ways to
reach an overall energy balance.
 Application: The applications themselves will determine the
location and the rate at which the nodes generate data. The area

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 11
generating more data and the path forwarding more packets may
suffer faster energy depletion.
 Routing: Most energy-efficient routing protocols always choose a
static optimal path to minimize energy consumption which results in
energy imbalance since the energy at the nodes on the optimal path
is quickly depleted.
5. SYSTEM DESIGN DESCRIPTION
5.1 EBERP: Energy Balanced with Efficiency Routing Protocol:
The goal of Energy Balanced with Efficiency Routing Protocol is to force
the packets to move towards the sink so that the nodes with relatively low
residual energy are protected. The Energy Balanced with Efficiency Routing
Protocol is designed by constructing a mixed virtual potential field. It forces
packets to move towards the sink through dense energy area. It protects the
sensor nodes with low residual energy. Successfully delivers the sensed
packet to the sink. Result shows significant improvement in network
lifetime, coverage ratio and throughput.
This article focuses on routing that balances the energy consumption with
efficiency. Its main contributions are:
 The concept of potential in classical physics is referred to build a
virtual hybrid potential field to drive packets to move towards the
sink through the high energy area and steer clear of the nodes with
low residual energy so that the energy is consumed as evenly as
possible in any given arbitrary network deployment.
 Classify the routing loops and devise an enhanced mechanism to
detect and eliminate loops. The simulation results reflect that the
proposed solution for EBERP makes significant improvements in
energy consumption balance, network lifetime and throughput when
compared to the other commonly used energy efficient routing
algorithm.
An energy-efficient routing protocol, tries to extend the network lifetime
through minimizing the energy consumption whereas energy balanced with
efficiency routing protocol intends to prolong the network lifetime through
uniform and efficient energy consumption. The former readily results in the
premature network partition that disables the network functioning, although
there may be much residual energy left. On the other hand, the latter may
not be optimal with respect to energy efficiency as it can burn energy evenly
to keep network connectivity and maintain network functioning as long as
possible. Let us use a simple example to demonstrate what uneven energy
depletion results in and how the proposed scheme Energy Balanced with

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 12
Efficiency Routing Protocol (EBERP) works to balance energy
consumption with efficiently.
In this system, energy balanced data gathering routing algorithm is
developed using the concepts of potential in classical physics. Our scheme
called energy balanced routing protocol, forwards data packets toward the
sink through dense energy areas so the nodes with relatively low residual
energy can be protected. The cornerstone of the EBERP is to construct three
independent virtual potential fields in terms of energy density, depth and
residual energy. The depth field is used to establish a basic routing paradigm
which helps in moving the packets towards the sink. The energy density
field ensures that packets are always forwarded along the high energy areas.
Finally, the residual energy field aims to protect the low energy nodes and
the energy is balanced efficiently.
5.2 Depth of Potential Field
To provide the basic routing function, namely to instruct packets move
toward the sink, we define the inverse proportional function of depth as the
depth potential field Vd(d) as shown in Eq. 5.1:
Where d =D (i) denotes the depth of node i. Then, the depth potential
difference Ud (d1; d2) from depth d1 to depth d2 is given by Eq 5.2
Since the potential function Vd(d) is monotonically decreasing, when the
packets in this depth potential field move along the direction of the gradient,
they could reach the sink eventually and the basic routing function can be
achieved. For a given network topology, Vd(d) is definite and time
invariant. Moreover, when the data packets move closer to the sink, the
centrality should be larger, where the centrality denotes the trend that a node
in depth d forwards the packets to the neighbors in depth d-1.
Figure 1. Depth potential field

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 13
 Energy Density Potential Field
A node adds up the energy values of all its neighbors, which can be obtained
through messages exchanged among nodes and calculates the area of the
radio coverage disk, so that the corresponding energy density can be readily
obtained using the aforementioned definition. EBERP defines the energy
density potential field as shown in Eq. 5.3 as follows:
Where Ved(i; t) is the energy density potential of node i at time t,
and ED(i; t) is the energy density on the position of node i at time t. Thus,
the potential difference Ued(i; j; t)from node i to node j is given by Eq. 5.4
Driven by this potential field, the data packets will always flow
toward the dense energy areas. However, with only this energy density field,
the routing algorithm is not practical since it would suffer from the serious
problem of routing loops. This fact will be clarified in the subsequent
simulation experiments.
 Energy Potential Field
EBERP defines an energy potential field as shown in Eq. 5.5 using the
residual energy on the nodes in order to protect the nodes with low energy:
Where Ve (i; t) is the energy potential of node i at time t, and E(i; t)
is the residual energy of node i at time t. Then potential difference Ue (i; j; t)
from node i to j is derived as shown in Eq 5.6.
The two latter potential fields are constructed using the linear
functions of energy density and residual energy, respectively. Although the
properties of the linear potential fields are straightforward, both of them are
time varying, which will result in the routing loop.
6. PERFORMANCE EVALUATION
In this section protocols are evaluated by simulation. It illustrates the
advantages of our protocol along with Mint Route protocol which uses the
shortest path for transfer of packets from source to sink.
6.1 Performance Metrics
To make a performance evaluation, several measurable metrics has
to be defined.
 Network Lifetime
The network lifetime [16] of a sensor network is defined as the time
when the first energy exhausted node (First Dead Node, FDN) appears. The
network lifetime is closely related to the network partition and network

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 14
coverage ratio. When a node begins to die, the probability of network
partition increases and the network coverage ratio might reduce.
 Functional lifetime
The functional lifetime of a task is defined as the amount of time
that the task is perfectly carried out. Different tasks have different
requirements. Some tasks may require no node failure while some others
just need a portion of nodes to be alive, therefore the function lifetime may
vary much according to task requirements. In simulation experiments,
requirements are based on the application by making all the sampling nodes
alive, functional lifetime is defined as the interval between the beginning of
task and the appearance of the First Dead Sampling Node (FDSN).
 Functional Throughput (FT)
Functional throughput is defined as the number of packets thatthe
sink receives during the functional lifetime. For a given application, FT is
mainly influenced by the length of the functional lifetime
6.2 Simulation Setup
The simulation experiments in wireless sensor networks are conducted and
evaluated to get the performance of our EBERP and compare them with
Mint Route algorithm. In this special topology, a node can only
communicate with its direct neighbors. The node can act as either a
sampling node or a relaying node depending on the requirements. The nodes
in the event areas can execute sampling and relaying tasks. The same
simulation is repeated by deploy in n number of nodes with a maximum of
1000 nodes, the average values of the performance metrics are calculated.
6.3 Performance Results.
In order to evaluate the relative performance of proposed protocol,
the protocol is compared with the existing Mint Route protocol. The graph
shown in the fig 3 will give a comparison result of how well the energy is
balanced for routing in our proposed scheme.
Figure 2. Comparison results for EBERP and Mint Route routing
 Network Lifetime and Network Throughput

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 15
Mint Route always chooses the shortest path, thus it will burn out the energy
of nodes on that path quickly. However, EBERP will choose another path
through other areas with more energy once it finds out that the energy
density in this area is lower than that in other areas nearby. Therefore,
EBERP can improve the energy consumption balance across the network
and prolong the network lifetime as well as the functional lifetime. The
statistical results are listed in table 8.1 shows the network throughput. The
EBERP prolongs the time of FDN. The functional throughput is and
network lifetime is also improved. The statistics listed in the table 8.2 show
the results of network lifetime. From these results, conclusion can be drawn
that more gain can be obtained through the EBERP’s energy consumption
balance and the integrity of the data received in EBERP is much better than
that in Mint Route since there is fewer packets loss in EBERP.
Figure 3. Network Throughput
Figure 4. Network Lifetime
6.4 Summary
The performance evaluation chapter discusses about the simulation results
drawn by considering all the performance metrics parameters like functional
lifetime, network lifetime and network throughput. The comparison
performance graph along with the network throughput and network lifetime
graph gives a clear overview of the existing and proposed protocols being
implemented.
7. CONCLUSION AND FUTURE ENHANCEMENT
7.1 Conclusion
Energy is an important resource for battery-powered wireless sensor
networks (WSN) that makes a key challenging problem for designing
energy-efficient protocol. Most of the existing energy efficient routing
protocols usually forward packets through the minimum energy path to the
0
1
2
1 2 3 4 5 6 7 8
Mint Routing-
Network
Throughput
0
0.5
1
74
147
220
293
366
439
512
585
658
731
X- axis Total
number of
packets sent

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 16
sink that merely minimizes energy consumption which leads unbalanced
distribution of residual energy amongst sensor nodes. Only, saving energy is
not enough to effectively prolong the network lifetime. The uneven energy
depletion often results in network partition and low coverage ratio which
decrease the performance. This article focuses on routing that balances the
energy consumption with efficiently. Its main contributions are firstly,
referring the concept of potential in classical physics to build a virtual
hybrid potential field to drive packets to move towards the sink through the
high energy area and steer clear of the nodes with low residual energy so
that the energy is consumed as evenly as possible in any given arbitrary
network deployment. Then, classify the routing loops and devise an
enhanced mechanism to detect and eliminate loops. The simulation results
reflect that the proposed solution for EBERP makes significant
improvements in energy consumption balance, network lifetime and
throughput when compared to the other commonly used energy efficient
routing algorithm.
7.2 Future Enhancement
In this project the routing loops: one hop - loop, origin - loop and queue -
loop are being detected and eliminated by cutting the loop. Hence, future
enhancement can be done in detecting and eliminating the loops and
transmitting packets by avoiding the loops. It will further help in improving
the overall system performance.
8. ACKNOWLEDGMENTS
This research would not have been possible without the help of my research
guide Dr. Mahadavan, Mr. Aziz Makandar, who gladly provided me with
the required information and equipment so that I could complete
myresearch. I would also like to thank our VC Dr. Meena R. Chandawarkar
who motivated me to take this work and for providing moral support.
REFERENCES
[1] Andrew S. Tanenbaum, Computer Networks, Prentice Hall of India Publications, 4th
Edition, 2006.
[2]Carlos Golmez, Joseph Padelles, Sensors Everywhere,Prentice Hall of India Publication,
4th Edition.
[3] J. Evans, D. Raychaudhuri, and S. Paul, “Overview of Wireless, Mobile and Sensor
Networks in GENI,” GENI Design Document 06- 14, Wireless Working Group, 2006.
[4] S. Olariu and I. Stojmenovi, “Design Guidelines for Maximizing Lifetime and Avoiding
Energy Holes in Sensor Networks with Uniform Distribution and Uniform Reporting,”
Proc. IEEE INFOCOM, 2006.
[5] Ian F. Akyildizon, Weilian Su, Yogesh Sankasubramanium and Erdal Cayirici, A
Survey on Sensor Networks, IEEE communications Magazine, August 2002, pp 102 – 114.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 17
[6] H. Zhang and H. Shen, “Balancing Energy Consumption to Maximize Network
Lifetime in Data-Gathering Sensor Networks,” IEEE Trans. Parallel and Distributed
Systems, vol. 20, no. 10, pp. 1526-1539, Oct. 2009.
[7] W. Heinzelman, A. Chandrakasan, and H. Balakrishnan, “Energy- Efficient
Communication Protocols for Wireless Microsensor Networks,” Proc. Hawaiian Int’l Conf.
Systems Science, 2000.
[8] O. Younis and S. Fahmy, “HEED: A Hybrid, Energy-Efficient Distributed Clustering
Approach for Ad Hoc Sensor Networks,” IEEE Trans. Mobile Computing, vol. 3, no. 4, pp.
366-379, Oct.-Dec. 2004.
[9] M. Singh and V. Prasanna, “Energy-Optimal and Energy-Balanced Sorting in a Single-
Hop Wireless Sensor Network,” Proc. First IEEE Int’l Conf. Pervasive Computing and
Comm., 2003.
[10] H. Lin, M. Lu, N. Milosavljevic, J. Gao, and L.J. Guibas, “Composable Information
Gradients in Wireless Sensor Networks,” Proc. Seventh Int’l Conf. Information Processing
in Sensor Networks (IPSN), pp. 121-132, 2008.
[11] Y. Xu, J. Heidemann, and D. Estrin, “Geography-Informed Energy Conservation for
Ad-Hoc Routing,” Proc. ACM MobiCom, 2001.
[12] V. Rodoplu and T.H. Meng, “Minimum Energy Mobile Wireless Networks,” IEEE J.
Selected Areas in Comm., vol. 17, no. 8, pp. 1333- 1344, Aug. 1999.
[13] W. Heinzelman, J. Kulik, and H. Balakrishnan, “Adaptive Protocols for Information
Dissemination in Wireless Sensor Networks,” Proc. ACM MobiCom, 1999.
[14] D.H. Armitage and S.J. Gardiner, Classical Potential Theory. Springer, 2001.
[15] C. Schurgers and M. Srivastava, “Energy Efficient Routing in Wireless Sensor
Networks,” Proc. Military Comm. Conf. (MILCOM), 2001.
[16]K. Kalpakis, K. Dasgupta, and P. Namjoshi, “Maximum Lifetime Data Gathering and
Aggregation in Wireless Sensor Networks,” Proc. IEEE Int’l Conf. Networking (ICN), pp.
685-696, 2002.
[17] AmolBakshi, Viktor K.Prasanna, “Energy-Efficient Communication in Multi-Channel
Single-Hop Sensor Networks.”Proceeding ICPADS ’04 Proceedings of the Parallel and
Distributed Systems, Tenth International Conference, page 403.
[18] Jing Wang , Dept. of ECE, North Carolina Univ., Charlotte, NC ,“Power Efficient
Stochastic - Wireless Sensor Networks.” Wireless Communications and Networking
Conference,2006.WCNC2006.IEEE, page 419-424.
[19] Li Hong , Shu-Ling Yang,“An Overall Energy-Balanced Routing Protocol for Wireless
Sensor Network.” Information and Automation for Sustainability, 2008.ICIAFS 2008. 4th
International Conference, page 314-318.
[20] Lin Wang, Ruihua Zhang, Shichao Geng,“An Energy-Balanced Ant-Based Routing
Protocol for Wireless Sensor Networks.” Wireless Communications, Networking and
Mobile Computing, 2009.WiCom '09. 5th International Conference on, page 1-4.
[21] Talooki, Marques. H, Rodriguez. J, Aqua, H. Blanco. N, Campos. L,“An Energy
Efficient Flat Routing Protocol for Wireless Ad Hoc Networks.” Computer
Communications and Networks (ICCCN), 2010 Proceedings of 19th International
Conference, page 1-6.
[22] http://en.wikipedia.org/wiki/Wireless_network.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 18
[23] http://oretan2011.wordpress.com/2011/01/28/wireless-sensor-network-wsn/
[24] http://en.wikipedia.org/wiki/WSN
Kawale, S. R., 2014. Enhancing Energy Efficiency in WSN using Energy Potential and Energy
Balancing Concepts. International Journal of Computer Science and Business
Informatics, Vol. 13, No. 1, pp. 9-18.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 19
DNS: Dynamic Network Selection
Scheme for Vertical Handover in
Heterogeneous Wireless Networks
M. Deva Priya
Department of CSE, Sri Krishna College of Technology,
Coimbatore, India.
D. Prithviraj
Department of CSE, Sri Krishna College of Technology,
Coimbatore, India.
Dr. M. L Valarmathi
Department of CSE, Government College of Technology,
Coimbatore, India.
ABSTRACT
Seamless Service delivery in a heterogeneous wireless network environment demands
selection of an optimal access network. Selecting a non-promising network, results in
higher costs and poor services. In heterogeneous networks, network selection schemes are
indispensable to ensure Quality of Service (QoS). The factors that have impact on network
selection include Throughput, Delay, Jitter, Cost and Signal Strength. In this paper, multi-
criteria analysis is done to select the access network. The proposed scheme involves two
schemes. In the first scheme, Dynamic Analytic Hierarchy Process (AHP) is applied to
dynamically decide the relative weights of the evaluative criteria set based on the user
preferences and service applications. The second scheme adopts Modified Grey Relational
Analysis (MGRA) to rank the network alternatives with faster and simpler implementation.
The proposed system yields better results in terms of Throughput, delay and Packet Loss
Ratio (PLR).
Keywords
Multi-Criteria Decision Making (MCDM) Scheme, Analytic Hierarchy Process (AHP),
Grey Relational Analysis (GRA), WiMAX, WiFi, QoS.
1. INTRODUCTION
Rapid development of multimedia applications in the wireless environment
has led to the development of many broadband wireless technologies. IEEE
802.16, a standard proposed by IEEE for Worldwide Interoperability for
Microwave Access (WiMAX) suggests modifications to the Medium
Access Control (MAC) and Physical (PHY) layers to efficiently handle high
bandwidth applications. IEEE 802.16 standards ensure Quality of Service

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 20
(QoS) for different types of applications supporting different types of
service classes[1].
1.1 IEEE 802.16 - WiMAX
IEEE 802.16, a solution to Broadband Wireless Access (BWA) is a wireless
broadband standard that promises high bandwidth over long range of
coverage[2]. The IEEE 802.16-2001 standard specified a frequency range
from 10 to 66 GHz with a theoretical maximum bandwidth of 120 Mbps and
a maximum transmission range of 50 kms. The initial standard supported
only the Line-Of-Sight (LOS) transmission and did not favor deployment in
urban areas.
IEEE 802.16a-2003 supports Non-LOS (NLOS) transmission and supports a
frequency range of 2 to11 GHz. IEEE 802.16 standard underwent several
amendments and evolved to the 802.16-2004standard (also known as
802.16d). It provided technical specifications to the PHY and MAC layers
for fixed wireless access and addresses the first or last mile connection in
Wireless Metropolitan Area Networks (WMANs).
IEEE 802.16e added mobility support. This is generally referred to as
mobile WiMAX and adds significant enhancements as listed below.
 It improves the NLOS coverage using advanced antenna diversity
schemes and Hybrid Automatic Repeat Request (HARQ).
 It adopts dense Subchannelization, thus increasing system gain and
improving indoor penetration.
 It uses Adaptive Antenna System (AAS) and Multiple Input Multiple
Output (MIMO) technologies to improve coverage.
 It introduces a DL Subchannelization scheme enabling better
coverage and capacity trade-off. This brings potential benefits in
terms of coverage, power consumption, self-installation and
frequency reuse and bandwidth efficiency.
With the rising popularity of multimedia applications in the Internet, IEEE
802.16 provides the capability to offer new wireless services such as
multimedia streaming, real-time surveillance, Voice over IP (VoIP) and
multimedia conferencing. Due to its long range and high bandwidth
transmission, IEEE 802.16 is also considered in areas where it can serve as
the backbone network with long separation among the infrastructure nodes.
Cellular technology using VoIP over WiMAX is another promising area.
WiMAX supports different types of traffics like Unsolicited Grant Service
(UGS), rtPS (real-time Polling Service), ertPS (extended real-time Polling
Service), nrtPS (non-real-time Polling Service) and Best Effort (BE).

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 21
Unsolicited Grant Service (UGS): Specifically designed for
Constant Bit Rate (CBR) services such as T1/E1 emulation and
VoIP without silence suppression.
Extended Real-Time Polling Service (ertPS): Built on the
efficiency of both the UGS and rtPS. This is suitable for applications
such as VoIP with silence suppression.
Real-Time Polling Service (rtPS): Designed for real-time services
that generate variable size data packets on periodic basis such as
MPEG video.
Non-Real-Time Polling Service (nrtPS): Designed for delay
tolerant services that generate variable size data packets on a regular
basis.
Best Effort (BE) Service: Designed for applications without any
QoS requirements such as HTTP service.
One of the main challenges in QoS provisioning is the effective mapping of
the QoS requirements of potential applications across different wireless
platforms [3].
1.1.1 Physical Layer
Orthogonal Frequency Division Multiplexing (OFDM) in the PHY layer
enables multiple accesses by assigning a subset of Subcarriers to users. This
resembles Code Division Multiple Access (CDMA) spread spectrum that
provides different QoS to each user. OFDM is achieved by multiplexing on
the user‟s data streams on both Uplink (UL) and Downlink (DL)
transmissions. The IEEE 802.16e Standard specifies the OFDMA based
PHY layer that has distinct features like flexible Subchannelization,
Adaptive Modulation and Coding (AMC), Space-time coding, Spatial
multiplexing, Dynamic Packet Switch based air interface and flexible
network deployment such as Fractional frequency reuse [7]. AMC
employed in the PHY layer dynamically adapts the modulation and coding
scheme to the channel conditions so as to achieve the highest spectral
efficiency at all times [8].
1.1.2 MAC Layer
The 802.16 MAC is designed to support a Point-to-Multipoint (PMP)
architecture with a central Base Station (BS) communicating simultaneously
with multiple Mobile Subscriber Stations (MSSs). The MAC includes the
following Sublayers namely:
Service Specific Convergence Sublayer (CS)- It maps the service
data units to the appropriate MAC connections, preserves or enables
QoS and bandwidth allocation.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 22
Common Part Sublayer (CPS)- It provides a mechanism for
requesting bandwidth, associating QoS and traffic parameters,
transporting and routing data to the appropriate convergence
Sublayer.
Privacy Sublayer - It provides authentication of network access and
assists in connection establishment [9].
1.2 IEEE 802.11 - WiFi
WLAN (or WiFi) is an open-standard technology that enables wireless
connectivity between equipments and Local Area Networks (LANs). Public
access WLAN services are designed to deliver LAN services over short
distances. Coverage extends over a 50 to 150 meter radius of the Access
Point (AP). Connection speeds range from 1.6 Mbps to 11 Mbps which is
comparable to fixed Digital Subscriber Line (DSL) transmission speed
[4].New standards promise to increase speeds upto 54 Mbps. Today‟s
WLANs run in the unlicensed 2.4 GHz and 5 GHz radio spectrums [5]. The
2.4 GHz frequency is already jam-packed - it is used for several purposes
besides WLAN service. The 5 GHz spectrum is a much larger bandwidth
providing higher speeds, greater reliability, and better throughput [6].
1.3 HANDOVER
Handover is the process of transferring an ongoing call or data session from
one channel connected to the core network to another. The WiMAX
technology specifies a variety of handover schemes to transfer a call or data
from the control of one network to another. When a MSS moves from one
BS to another, the control information is transferred from the BS to which
the MSS is currently linked referred to as the home Base Station (hBS) to
the BS under the range of which the MSS is to be connected referred to as
target Base Station (tBS).
Handover is of two types based on the technology of the networks involved
namely, Horizontal Handover and Vertical Handover. Figure. 1 illustrates
the WiMAX - WiFi network architecture where the MSS is handed over to
the optimal nearby BS or AP. The handovers based on access networks
include:
Horizontal Handover-The mobile user switches between networks
with the same technology.
Vertical Handover (VHO) -The users switch among networks with
different technologies, for example, between an IEEE 802.11 AP and
a cellular network BS. In heterogeneous networks, VHO is mainly
used. Users can move between different access networks. They
benefit from different network characteristics (coverage, bandwidth,
frequency of operation, data rate, latency, power consumption, cost,
etc.) that cannot be compared directly [10].

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 23
Figure 1. WiMAX - WiFi Network Architecture
2. RELATED WORK
A link reward function and a signaling cost function are presented in [11] to
capture the tradeoff between the network resources utilized by the
connection and the signaling and processing load acquired on the network.
A stationary deterministic policy is obtained when the connection
termination time is geometrically distributed.
A novel optimization utility is presented in [12] to assimilate the QoS
dynamics of the available networks along with heterogeneous attributes of
each user. The joint network and user selection is modelled by an
evolutionary game theoretical approach and replicator dynamics is figured
out to pursue an optimal stable solution by combining both self-control of
users‟ preferences and self-adjustment of networks‟ parameters.
A survey on fundamental aspects of network selection process is discussed
in [13]. It deals with network selection to the always best connected and
served paradigm in heterogeneous wireless environment as a perspective
approach.
A mechanism [14] based on a unique decision process that uses
compensatory and non-compensatory multi-attribute decision making
algorithms is proposed, which jointly assists the terminal in selecting the top
candidate network.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 24
A cross layer architectural framework for network and channel selection in a
Heterogeneous Cognitive Wireless Network (HCWN) is proposed in [15]. A
novel probabilistic model for channel classification based on its adjacent
channels‟ occupancy within the spectrum of an operating network is also
introduced. Further, a modified Hungarian algorithm is implemented for
channel and network selection among secondary users.
In [16], a Satisfaction Degree Function (SDF) is proposed to evaluate the
available networks and find the one that can satisfy the mobile user. This
function not only considers the specific network conditions (e.g. bandwidth)
but also the user defined policies and dynamic requirements of active
applications.
In [17], a two-step vertical handoff decision algorithm based on dynamic
weight compensation is proposed. It adopts a filtering mechanism to reduce
the system cost. It improves the conventional algorithm by dynamic weight
compensation and consistency adjustment.
A speed-adaptive system discovery scheme suggested in [18] for execution
before vertical handoff decision improves the update rate of the candidate
network set. A vertical handoff decision algorithm based on fuzzy logic
with a pre-handoff decision method which reduces unnecessary handoffs,
balancing the whole network resources and decreasing the probability of call
blocking and dropping is also added.
In [19], the authors present a multi-criteria vertical handoff decision
algorithm for heterogeneous wireless networks based on fuzzy extension of
TOPSIS. It is used to prioritize all the available networks within the
coverage of the mobile user. It achieves seamless mobility while
maximizing end-users' satisfaction.
A network selection mechanism based on two Multi Attribute Decision
Making (MADM) methods namely Multiple - Analytic Hierarchy Process
(M-AHP) and Grey Relational Analysis (GRA) method is proposed in [20].
M-AHP is used to weigh each criterion and GRA is used to rank the
alternatives.
A context-aware service adaptation mechanism is presented for ubiquitous
network which relies on user-to-object, space-time interaction patterns
which helps to perform service adaptation [21]. Similar Users based Service
Adaptation algorithm (SUSA) is proposed which combines both Entropy
theory and Fuzzy AHP algorithm (FAHP).
Load balancing algorithm based on AHP proposed in [22] helps the
heterogeneous WLAN/UMTS network to provide better service to high
priority users without decreasing system revenue.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 25
3. CROSS LAYER DESIGN
To ensure seamless QoS, a Cross-Layered Framework is designed for
network selection in heterogeneous environments. The PHY layer, MAC
(L2) layer and the Network layer ((L3) are involved. The layers are closely
coupled together (Figure 2).
TIER-1: It includes the PHY and the MAC layers. Resource
availability is determined from the MAC layer. The parameters RSSI
and SINR are taken from the PHY layer.
TIER-2: In the Network layer, network is selected for a MSS based
on the factors determined from TIER-1.
Figure 2. Cross Layer Design
4. MULTI- CRITERIA DECISION MAKING (MCDM) SHEMES
Handover decision problem deals with selecting network from candidate
networks of various service providers involving technologies with different
criteria. Network selection schemes can be categorized into two types -
Fuzzy Logic based schemes and Multiple Criteria Decision Making
(MCDM) based schemes.
Three different approaches for optimal access network selection are [23,
24]:
Network Centric - In network centric approach, the choice for
access network selection is made at the network side with the goal of
improving network operator‟s benefit. Majority of network centric
approaches use game theory for network selection.
User Centric - In this approach, the decision is taken at the user
terminal based only on the minimization of the user‟s cost without
considering the network load or other users. The selection of the
access network is determined by using utility, cost or profit functions

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 26
or by applying MCDM methods. The selection of an access network
depends on several parameters with different relative importance
such as network and application characteristics, user preferences,
service and cost.
Collaborative Approaches - In the collaborative approach,
selection of access network takes into account the profits of both the
users and the network operator. It mainly deals with the problem of
selecting a network from a set of alternatives which are categorized
in terms of their attributes.
The two processes in MCDM techniques are weighting and ranking. Most
popular classical algorithms include Simple Additive Weighting (SAW),
Technique for Order Preference by Similarity to Ideal Solution (TOPSIS),
Analytical Hierarchy Process (AHP) and Grey Relational Analysis (GRA).
 In Simple Additive Weighting (SAW), the overall score of a
candidate network is determined by the weighting sum of all the
attribute values.
 In Technique for Order Preference by Similarity to Ideal Solution
(TOPSIS), the chosen candidate network is one which is closest to
the ideal solution and farthest from the worst case solution.
 Analytical Hierarchy Process (AHP) decomposes the network
selection problem into several subproblems and assigns a weight for
each subproblem.
 Grey Relational Analysis (GRA) ranks the candidate networks and
selects the one with the highest ranking.
5. ANALYTIC HIERARCHY PROCESS (AHP)
AHP was introduced by Saaty [25] with the goal of making decisions about
complex problems by dividing them into a hierarchy of decision factors
which are simple and easy to analyze.
 AHP generates a weight for each evaluation criterion according to
the decision maker‟s pairwise comparisons of the criteria. The higher
the weight, the more important the corresponding criterion.
 Next, for a fixed criterion, it assigns a score to each option according
to the decision maker‟s pairwise comparisons of the options based
on that criterion. The higher the score, the better the performance of
the option with respect to the considered criterion.
 Finally, the AHP combines the criteria weights and the options
scores thus determining a global score for each option and a
consequent ranking. The global score for a given option is the
weighted sum of the scores obtained with respect to all the criteria.
6. DYNAMIC ANALYTIC HIERARCHY PROCESS (DAHP)
In the proposed Dynamic AHP (DAHP), the weight of each criterion is
assigned dynamically based on the Received Signal Strength Indicator

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 27
(RSSI) and Signal to Noise Interference Ratio (SINR) values of a MSS with
respect to a BS or AP. A network with high RSSI and low SINR is given
priority. Likewise, the values of both RSSI and SINR are calculated at
regular intervals and the weights are assigned. Table 1 shows the possible
weights that are assigned to a network based on the parameter values.
Table 1: Weights Assignment based on values
DAHP involves the following steps:
Step 1: Determination of the objective and the decision factors:
In this step, the final objective of the problem is analyzed based on a
number of decision factors. They are further analyzed until the
problem acquires a hierarchical structure. In the lowest level, the
alternative solutions of the problem are found (Figure 3).
Step 2: Determination of the relative importance of the decision
factors with respect to the objective: In each level, decision factors
are pairwise compared according to their levels of influence with
respect to the scale in Table 1. If there are „n‟ decision factors, then
the total number of comparisons will be „n (n - 1)/2‟. For qualitative
data such as preference, ranking and subjective opinions, it is
suggested to use a scale from 1 to 7 as shown in Table 2.
Table 2: Scale of Importance
PREFERENCE LEVELS VALUES
Equally preferred 1
Equally to moderately preferred 2
Moderately preferred 3
Moderately to strongly preferred 4
Strongly preferred 5
Strongly to very strongly preferred 6
Very strongly preferred 7
RESOURCE
AVAILABILITY
RSSI SINR
SELECT/R
EJECT
AVAILABLE
High High
Select (Worst
Case)
High Medium Select
High Low Select
Medium High Reject
Medium Medium Select
Medium Low Select
Low High Reject
Low Medium Reject
Low Low Reject

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 28
Figure 3. Hierarchy of criteria and alternatives
Initially, a pair-wise comparison „n×n‟ matrix „A[i][j]‟ is formed, where „n‟
is the number of evaluation criterion considered. Each entry „aij ‟ of the
matrix represents the importance of the criterion relative to the „ jth
‟
criterion.
If aij=1, an element is compared with itself.
If aij>1,then element „i‟ is considered to be more important than
element „j‟.
If aij<1,then element „j‟ is considered to be more important than
element „i‟.
aij =
1
aji
for the rest of the values of the table.
Each entry is multiplied with the respective parameter values which
increases the accuracy of the criterion weights.
The entries „ajk ‟ and „akj ‟ satisfies the following constraint:
ajk ∗ akj = 1 (1)
Also,ajj = 1 for all „j‟.
Step3: Normalization and calculation of the relative weights:
Relative weight is a ratio scale that can be divided among decision
factors. The relative weights are calculated by following the steps
given below.
 Each column of matrix A is summed.
 Each element of the matrix is divided by the sum of its column.
The relative weights are normalized. After normalizing, the sum
of each column is one.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 29
 Normalized principle Eigen vector is obtained by finding the
average of rows after normalizing.
 A priority vector is obtained which shows the relative weights
among decision factors that are compared. Normalized principle
Eigen vector gives the relative ranking of the criteria used.
 For consistency, largest Eigen value (λmax) is obtained from the
summation product of each element of the Eigen vector and sum
of columns of matrix A.
When many pairwise comparisons are performed, some inconsistencies
typically arise. AHP incorporates an effective technique for checking the
consistency of the evaluations made by the decision maker when building
each pairwise comparison matrix involved in the process and it mainly
depends on the computation of a suitable Consistency Index (CI). The CI is
obtained by computing the scalar „x‟ as the average of the elements of the
vector whose „jth
‟ element is the ratio of the „jth
‟ element of the vector
„A*w‟ to the corresponding element of the vector „w‟.
CI =
λmax − n
n−1
(2)
A perfectly consistent decision maker should always yield CI=0. Small
values of inconsistency may be tolerated. RI is the Random Index, i.e. the
CI when the entries of „A‟ are completely random. The values of RI for
small problems (m ≤ 10) are shown in Table 3.
Table 3: Values for Random Index
In particular, if
CI
RI
≤10%, the inconsistency is acceptable and a reliable
result may be expected. If the consistency ratio is greater than 10%, pairwise
comparison should be initiated from the beginning.
7. MODIFIED GREY RELATIONAL ANALYSIS (MGRA)
Grey system theory is one of the methods used to study uncertainty and is
considered superior in the mathematical analysis of systems with uncertain
information. A system with partial information is called a grey system. GRA
is a part of grey system theory which is suitable for solving problems with
complicated interrelationships between multiple factors and variables. GRA
method is widely used to solve the uncertainty problems with discrete data
1 2 3 4 5 6 7 8 9 10
0 0 0.58 0.9 1.12 1.24 1.32 1.41 1.45 1.49

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 30
and incomplete information. One of the sequences is defined as reference
sequence presenting the ideal solution. The grey relationship between the
reference sequence and other sequences can be determined by calculating
the Grey Relational Coefficient (GRC). MGRA involves the following
steps.
Step 1: Classifying the series of elements into three categories:
larger-the-better, smaller-the-better and nominal-the-best.
Step 2: Defining the lower, moderate or upper bounds of series
elements and normalizing the entities.
Step 3: Calculating the GRCs.
Step 4: Selecting the alternative with the largest GRC.
The upper bound (uj) is defined as
max{S1(j), S2(j), …, Sn(j)} (3)
and the lower bound (lj) is calculated as
min{S1(j), S2(j), …, Sn(j)},(4)
For the moderate bound (mj), the objective value between the lower and
upper bound is considered.
 The absolute difference between „Si(j)‟ and „lj‟ or „uj‟ divided by the
difference between „lj‟ and „uj‟ achieves the normalization „Si
∗
j ‟
for larger or smaller, where i = 1… n.
 The normalization for nominal-the-best is presented as „uj‟ for
larger-the-better, „lj‟ for smaller-the-better and „mj‟ for nominal-the-
best. They are chosen to form a reference series „S0‟ which actually
presents the ideal situation.
The GRC is computed from
GRCi =
1
wj Si
∗ j −1k
j=1 +1
(5)
where wj is the Weight of each parameter.
The comparative series with the largest GRC is given the highest priority.
8. RESULTS AND DISCUSSION
A heterogeneous network scenario is simulated using ns2. The simulation
parameters are shown in Table4.Three different types of SLAs namely
SLA1 (High), SLA2 (Medium) and SLA3 (Low) are considered.
 The most important selection criterion for SLA1 is the QoS
satisfaction degree and not the cost of service.
 On the other hand, Cost criterion is more important than the degree
of perceived QoS for SLA2 and SLA3.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 31
When a Service Provider does not have resources or the QoS is not good,
the users are moved to a WiFi network to improve the performance.
Table 4: Simulation Parameters
PARAMETER VALUE
MAC Mac/802.16e & 802.11
Packet Size 5000
Bandwidth 1 Mbps
Queue Length 50
Routing DSDV
Simulation time 50 Sec
The Throughput (Figure 4)of the proposed DAHP is better when compared
to the existing scheme. The proposed scheme offers 1.15, 1.11 and 1.05
times more Throughput when compared to AHP for SLA1, SLA2 and SLA3
respectively.
Figure 4. Throughput
The proposed scheme offers 1.03, 1.2 and 1.1 times less cost when
compared to AHP for SLA1, SLA2 and SLA3 respectively (Figure 5).
Figure 5. Cost
The Average Delay (Figure 6) of the AHP scheme is 1.46, 1.38 and 1.2
times more than that of DAHP.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 32
Figure 6. Delay
The proposed scheme offers 1.26, 1.19 and 1.24 times less Average Jitter
when compared to AHP for SLA1, SLA2 and SLA3 respectively (Figure 7).
Figure 7. Jitter
Similarly, the Packet Loss Ratio (PLR) of DAHP is less when compared to
former scheme as network selection is done dynamically based on the QoS
values (Figure 8). The PLR of AHP scheme is 1.21, 1.12 and 1.13 times
more than that of DAHP.
Figure 8. Packet Loss Ratio
9. CONCLUSION
An optimal network selection scheme is proposed for heterogeneous
networks. The physical layer parameters such as Signal Strength and Noise

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 33
Ratio are integrated. This scheme dynamically weighs every possible
candidate network for MSSs using DAHP and each is ranked by the MGRA.
The proposed network selection algorithm provides seamless connection for
the users over the heterogeneous network and enables the MSSs to forward
the calls to the optimal network without dropping it. The simulation results
reveal that the proposed network selection scheme efficiently decides the
trade-off among user preference and network condition. It offers better
Throughput involving less Cost, Delay, Jitter and PLR. In the future, the
proposed scheme can be enhanced to include more network alternatives and
selection criteria.
REFERENCES
[1] Haghani, E., Parekh, S., Calin, D., Kim, E. and Ansari, N. “A quality-driven cross-
layer solution for MPEG video streaming over WiMAX networks”, IEEE
Transactions on Multimedia, Vol. 11, No. 6, pp. 1140-1147, 2009.
[2] IEEE Std 802.16-2009, “IEEE standard for local and metropolitan area networks”,
Part 16: Air interface for broadband wireless access systems, 2009.
[3] Bo Li, Yung Qin, Chor Ping Low and Choon Lim Guee. “A survey on mobile
WiMAX”, IEEE communications magazine, pp. 70-75, 2007.
[4] Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY)
Specification, IEEE 802.11 WG, Aug. 1999.
Specification: High-Speed Physical Layer Extension in the 2.4 GHz Band, IEEE
802.11b WG, Sept. 1999.
Specification: High-Speed Physical Layer in the 5 GHz Band, IEEE 802.11a WG,
Sept. 1999.
[7] Kennington, J., Olinick, E. and Rajan, D. “Wireless network design - optimization
models and solution procedures”, Springer, 2010.
[8] Ali-Yahiya, T., Beylot, A. and Pujolle, G. “An adaptive cross-layer design for
multiservice scheduling in OFDMA based mobile WiMAX systems”, Computer
Communications, Vol. 32, pp. 531-539, 2009.
[9] Eklund, C., Marks, R.B., Stanwood, K.L. and Wang, S. “IEEE Standard 802.16: A
technical overview of the Wireless MAN™ air interface for broadband wireless
access”, IEEE Communications Magazine, pp. 98-107, 2002.
[10] Nasser, N., Hasswa, A., and Hassanein, H. “Handoffs in fourth generation
heterogeneous networks”, IEEE Communications Magazine, Vol. 44, pp.96-103, 2006.
[11] Stevens-Navarro, E., Lin, Y. and Wong, V. W. “An MDP-based vertical handoff
decision algorithm for heterogeneous wireless networks”, IEEE Transactions on
Vehicular Technology, Vol. 57, No. 2, pp. 1243-1254, 2008.
[12] Pervaiz, Haris, Qiang Ni, and Charilaos C. Zarakovitis. “User adaptive QoS aware
selection method for cooperative heterogeneous wireless systems: A dynamic
contextual approach”, Future Generation Computer Systems, 2014.
[13] Rao, K. R., Zoran S. Bojkovic, and Bojan M. Bakmaz. “Network selection in
heterogeneous environment: A step toward always best connected and served”, In 11th
International Conference on Telecommunication in Modern Satellite, Cable and
Broadcasting Services (TELSIKS), Vol. 1, pp. 83 - 92, 2013.
[14] Bari, F.and Leung, V. C. “Automated network selection in a heterogeneous wireless
network environment”, IEEE Network, Vol. 21, No. 1, pp. 34-40, 2007.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 34
[15] Haldar, Kuheli Louha, Chittabrata Ghosh, and Dharma P. Agrawal. “Dynamic
spectrum access and network selection in heterogeneous cognitive wireless
networks”, Pervasive and Mobile Computing, Vol. 9, No. 4, pp. 484 - 497, 2013.
[16] Cai, X., Chen, L., Sofia, R., & Wu, Y. “Dynamic and user-centric network selection in
heterogeneous networks”, In IEEE International Performance, Computing, and
Communications Conference (IPCCC), pp. 538-544, 2007.
[17] Liu, Chao, Yong Sun, Peng Yang, Zhen Liu, Haijun Zhang, and Xiangming Wen. “A
two-step vertical handoff decision algorithm based on dynamic weight compensation”,
In International Conference on Communications Workshops (ICC), pp. 1031 - 1035,
2013.
[18] Yang, Peng, Yong Sun, Chao Liu, Wei Li, and Xiangming Wen, “A novel fuzzy logic
based vertical handoff decision algorithm for heterogeneous wireless networks”,
In 16th
International Symposium on Wireless Personal Multimedia Communications
(WPMC), pp. 1 - 5, 2013.
[19] Mehbodniya, Abolfazl, Faisal Kaleem, Kang K. Yen, and Fumiyuki Adachi. “A novel
wireless network access selection scheme for heterogeneous multimedia traffic”,
In Consumer Communications and Networking Conference (CCNC), pp. 485- 489,
2013.
[20] Lahby, Mohamed, and Abdellah Adib. “Network selection mechanism by using M-
AHP/GRA for heterogeneous networks”, In 6th
Joint IFIP Wireless and Mobile
Networking Conference (WMNC), pp. 1-6, 2013.
[21] Chang, Jie, and Junde Song. “Research on Context-Awareness Service Adaptation
Mechanism in IMS under Ubiquitous Network”, In 75th
Vehicular Technology
Conference (VTC Spring), pp. 1-5, 2012.
[22] Song, Qingyang, Jianhua Zhuang, and Rui Wen. “Load Balancing in WLAN/UMTS
Integrated Systems Using Analytic Hierarchy Process”, In Recent Advances in
Computer Science and Information Engineering, Springer Berlin Heidelberg, pp. 457-
464, 2012.
[23] Hwang, C. L., and Yoon, K. “Multiple attribute decision making: Methods and
applications”, in A state of the art survey, New York: Springer, 1981.
[24] Meriem, K., Brigitte, K., and Guy, P. “An overview of vertical handover decision
strategies in heterogeneous wireless networks”, Journal of Computer, Communication,
Elsevier, Vol. 37, No. 10, 2008.
[25] Saaty, T. L. The analytical hierarchy process, planning, priority setting, resource
allocation, NewYork: Mcgraw Hill, 1980.
Priya, M. D., Prithviraj, D. and Valarmathi, M. L., 2014. DNS: Dynamic
Network Selection Scheme for Vertical Handover in Heterogeneous
Wireless Networks. International Journal of Computer Science and
Business Informatics, Vol. 13, No. 1, pp. 19-34.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 35
Implementation of Image based
Flower Classification System
Tanvi Kulkarni
PG Student
Department of IT, SCOE, Pune
Nilesh. J. Uke
Associate Professor
Department of IT, SCOE, Pune
ABSTRACT
In today’s world, automatic recognition of flowers using computer technology is of great
social benefits. Classification of flowers has various applications such as floriculture,
flower searching for patent analysis and much more. Floriculture industry consists of flower
trade, nursery and potted plants, seed and bulb production, micro propagation and
extraction of essential oil from flowers. For all the above, automation of flower
classification is very essential step. However, classifying flowers is not an easy task due to
difficulties such as deformations of petals, inter and intra class variability, illumination and
many more. The flower classification system proposed in this paper uses a novel concept of
developing visual vocabulary for simplifying the complex task of classifying flower
images. Separate vocabularies for color, shape and texture features are created and then
they are combined into final classifier. In this process firstly, an image is segmented using
grabcut method. Secondly, features are extracted using appropriate algorithms such as SIFT
descriptors for shape, HSV model for color and MR8filter bank for texture extraction.
Finally, the classification is done with multiboost classifier. Results are represented on 17
categories of flower species and seem to have efficient performance.
Keywords
MR8 filter bank, Multiboost classifier, SIFT descriptors, Visual Vocabulary, HSV color
model.
1. INTRODUCTION
Object recognition has always been a difficult problem to tackle for the
computer scientists due to the numerous challenges involved in it. It is
possible that the image of any object taken from different view appears in a
different way for each individual. Considering the natural object such as
flower, various species of flowers exists in the world. Some of the
categories are Daffodils, Buttercups, Dasils, Iris, Dandelions, Paisy,
Sunflowers, Windflowers, Lily valleys, Tulips, Tiger lilies, Crocus,
Bluebells, Cow clips etc. The categorization of flower images is challenging
due to variances in geometry, illumination and occlusions. The problem of
classification becomes more complex because of the large visual variation

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 36
between images of same flower species known as inter-class variability and
variation between images of different flower species called intra-class
variability. Figure1 depicts the three different kinds of flowers having
similar shape and appearance thus showing the inter-class variability.
Figure1. Flower images for inter class variability
Hence, there is a need to create a classification system that captures the
important aspects of a flower and also address issues such as variation in
illumination, occlusion, view angle, rotation and scale. This paper focuses
on proposing a system that can classify flower images by developing a
visual vocabulary that represents different distinguishing aspects of flower.
This system thus can overcome ambiguities that exist between flower
categories.
The rest paper is organized as follows: Section 2 briefs about the work done
till now related to this area. The implementation of flower classification
system using visual vocabulary is discussed in Section 3. Results of various
techniques implemented are discussed in section 4. Section 5 concludes this
paper.
2. RELATED WORKS
Many researchers have worked on the various methods and algorithms for
the flower image classification. Nilsback and Zisserman have proposed a
novel concept of visual vocabulary in order to address the issue of
ambiguity [5]. Wenjing Qi, et al. has suggested the idea of flower
classification based on local and spatial cues with help of SIFT feature
descriptors [8]. Yong Pei and Weiqun Cao has provided the application of
neural network for performing digital image processing for understanding
the features of a flower [10].Regional feature extraction method based on
shape characteristics of flower is proposed by Anxiung Hong, Zheru Chi, et
al.[7]. Salahuddin et al. have proposed an efficient segmentation method
which combines color clustering and domain knowledge for extracting
flower regions from flower images [4]. D S Guru et al. have developed an
algorithmic model for automatic flowers classification using KNN as the
classifier [3]. Nilsback and Zisserman has also computed four different
features for the flowers, each describing different aspects such as the local
shape/texture, the shape of the boundary, the overall spatial distribution of
petals, and the color. Finally they combined the features using a multiple
kernel framework with a SVM classifier [6].

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 37
4. METHODOLOGY
Recently, bag of visual words model [1] has gained tremendous success in
object classification. Visual Vocabulary [5] concept is based on the same
model. The most distinguishing characteristics of a flower image are the
shape, color and texture. Based on these features it becomes easy to classify
the flower images. Since the system is based on the concept of visual
vocabularies, separate vocabularies are created for color, shape and texture
features and the results are combined into final classifier. Detailed
description about the flow of the system is depicted in Figure.3.The entire
system works in two phases:-the training phase and secondly the testing
phase.
Figure 2. Block diagram of flower classification system
In training phase, all the images from all classes are selected and then their
color, shape and texture features are extracted with their respective
extraction techniques which are discussed later. The outcomes of this are the
descriptors which are provided as an input to k-means clustering algorithm
in order to form visual words. Using visual words, object histogram are
created .These histogram are given to the final multiboost classifier in order
to train them. In testing phase, when user provides the query image, firstly
feature extraction is performed then object histogram is created and given to
the classifier which with the help of trained parameters classifies the image
and provides it with the appropriate label.
The implementation of Visual vocabulary is explained below:-

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 38
1. SEGMENTATION-
The flower images that are taken from the dataset should be segmented first
in order to achieve the higher rate of accuracy. In this system, grabcut
method is used for segmentation and it yields good results. Grabcut is a
segmentation technique that uses region and boundary information in order
to perform segmentation. This information is gained through significant
difference between the colors of nearby pixels.
(a)Original image (b)Segmented image
Figure 3. Segmentation with grabcut method
Above figure (a) depicts the input flower image randomly selected from
dataset.Figure (b)shows the result of segmented flower image through the
grabcut method.
2. CREATING A VOCABULARY FOR FLOWER-
In order to create a flower vocabulary, we need to extract the feature
descriptors from the flower images using relevant methods and create
vocabularies of those.
A. SHAPE VOCABULARY-
Shape is the most important characteristic of flower. However, the natural
deformations of flowers and the variations of viewpoint and occlusions
change the original shape of the flower. To create rotation and scale
invariant shape descriptors, SIFT (Scale Invariant Feature Transform)
descriptors are the best method so they are extracted from flower images
which forms 128 dimensional vector. SIFT descriptors found in all training
images are clustered to create shape visual words.
Figure 4. SIFT keypoints extraction
To represent an image, a histogram is created based on the distance between
the observed SIFT descriptors [18] in the image and the computed cluster
centers. Figure 4 shows the keypoints calculated for the shape feature
extraction of a segmented flower image.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 39
B. COLOR VOCABULARY-
Color helps us to simplify the task of categorization. The effect of varying
illumination has an adverse effect on the measured color, which may lead to
confusion. HSV (Hue, Saturation and Value) color model hence is the most
efficient way of describing color. HSV color space is less sensitive to
illumination variations. Color visual words are created by clustering the
HSV value of each pixel in the training images. The computed cluster
centers represent the color visual words which comprises the color
vocabulary.
C.TEXTURE VOCABULARY-
Flowers can have distinctive or subtle textures on their petals. The texture is
described by convolving the images with filters from an MR8 (Maximum
Response) filter bank which is rotational invariant.MR filter bank generally
contains 38 filters. An MR8 filter consists of an edge and a bar filter at six
orientations and three scales, and two rotationally symmetric filters.
Figure 5. Convolving images with the MR8 filters
The 38 responses are summarized into eight maximum responses (three
scales for edge and bar filters, one each for Gaussian and Laplacian of
Gaussian).Figure 6 describes the results after convolving segmented image
with the MR8 filter bank.
D.COMBINED VOCABULARY-
The discriminative power of color, shape and texture varies for different
flower species. Some flowers can be more easily distinguished by their
shape, color and texture. However, it is better that, flowers are distinguished
by combination of these aspects. In order to distinguish a flower by these 3
aspects, they are combined in the classification system. They are combined
by assigning weights to their separate classification and not averaging them.
The Multiboost classifier [2] is used as it reduces variance and is less
sensitive to noise. Multiboost is an implementation of an extension of the
multi-class Adaboost algorithm.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 40
5. RESULTS
Considering the overall flower classification system,some of the
implementation results are discussed below.Firstly, when an input image is
selected for the categorisation purpose it is necessary that the image is
segmented.Following figure depicts the result shown by grabcut
segmentation method.
(a) (b) (c)
Figure 6. Segmentatation with grabcut method
Figure.(a) shows an input image randomly selected from
database.Fig.(b)shows segmented image through the grabcut technique
wherein background part is represented by black pixels and foreground part
by white pixels.Finally the white pixels are replaced by original color pixels
which is shown in fig.(c).It is the final segmented image of flower which is
to be used for further processing.
After segmentation,next step is feature extraction.First is the shape feature
extraction done through SIFT descriptors.Below figure descibes how
keypoints are calculated and stored.
Figure 7. SIFT keypoints detection
For the above flower image the numbers of keypoints calculated are: 65.
HSV color model is used for color feature extraction. The figure given
below is the HSV representation of original segmented flower image.
Figure 8. HSV color map

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 41
Finally, the texture feature is extracted by MR8 filter bank. The result after
convolving a segmented flower image with MR8 filters is described in
below figure.
Figure 9. Result of convolving image with MR8 filter bank
After the feature extraction process, bag of visual words will be created by
k-means clustering. Based on visual words histograms will be created and
provided to multiboost algorithm for training and then finally testing will be
performed through the query image from the user.
Considering single feature, classification does not prove to be as efficient as
by combining the three features together.12 images are considered as
training images and 3 images are taken for testing purpose. Below shows the
classification of flowers based on single feature. Whole data set is divided
into training and testing set for better classification purpose.
1. Classification based on Color feature-
It is sometimes not possible to classify the flower image just on the basis of
color .It is possible that two flowers have same color. For instance say,
daffodils and dandelions have same color yellow. For our classification
system when LilyValley was given as a query image the classified image
was of Snowdrop just purely based on white color.
Figure 10. LilyValley classified as Snowdrop based on white color
2. Classification based on Shape feature-
Shape helps to narrow down the flower species. Given a test image of
daffodils it was classified as daffodils only.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 42
Figure 11. Daffodils classified as Daffodils based on shape
3. Classification based on Texture feature-
Texture feature helps to improve the classification efficiency of a flower
image. When LilyValley was given as input result was the Snowdrop based
on the pattern.
Figure 12. LilyValley classified as Snowdrop based on texture
4. Classification based on combined feature-
Since it is not sufficient to classify flower images based on single feature
only, categorization based on combined features helps to improve the
performance of classification.
Figure 13. Daffodils classified as Daffodils based on combined
(Color+Shape+Texture) features
If we consider the classification based on individual features, accuracy for
each is described in the following graph. Highest accuracy of shape feature
is achieved of 77.27% with 25 folds.
Color feature achieves the accuracy of 85.50% with 20 folds. Texture
feature achieves the highest accuracy with 25 folds of 72.29%.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 43
Figure 14. Performace analysis of Shape, Color and Texture features
Considering the low efficiency of classification based on only the individual
features, combined features with multiboost classifier provides the best
results. Performance accuracy of85.98% is achieved with the combined
features.
Figure 15. Performace analysis of Combined (Shape, Color and Texture) features
6. CONCLUSION
Flower classification is slowly becoming the popular area owing to its
importance for botanists and in floriculture. Flower classification system
which is discussed in this paper will provide efficient classification accuracy
owing to the idea of visual vocabulary. Developing and combining
vocabularies for several aspects (color, shape and texture) of a flower image
boost the performance significantly. Moreover the final classifier adds to the
superiority of the performance. Thus, the tedious task of classifying various
flower images into appropriate categories is simplified in effective manner.
Performance analysis shows that combining features into final classifier
boosts the performance of flower classification rather than classifying based
on individual features.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 44
REFERENCES
[1] Csurka, Gabriella, et al.Visual categorization with bags of keypoints. Workshop on
statistical learning in computer vision, ECCV. Vol. 1. 2004.
[2] Freund, Y., Schapire, R.A decision-theoretic generalization of on-line learning and an
application to boosting.EuroCOLT”95 Proceedings of the Second European
Conference on Computational Learning Theory, pp. 23-37, 1995.
[3] Guru, D. S., Y. H. Sharath, and S. Manjunath. Texture features and KNN in
classification of flower images.IJCA, Special Issue on RTIPPR (1) (2010): 21-29,
2010.
[4] Hong, Anxiang, et al. Region-of-Interest based flower images retrieval. Acoustics,
Speech, and Signal Processing.2003 Proceedings. (ICASSP'03) IEEE International
Conference on. Vol. 3, 2003.
[5] Nilsback and Andrew Zisserman.A Visual Vocabulary for Flower Classification.
Computer Vision and Pattern Recognition, IEEE Computer Society Conference on.
Vol.2, 2006.
[6] Nilsback, M-E., and Andrew Zisserman. Automated flower classification over a large
number of classes. Computer Vision, Graphics & Image Processing, 2008.
[7] Pei, Yong, and Weiqun Cao. A method for regional feature extraction of flower
images.Intelligent Control and Information Processing (ICICIP), IEEE, 2010.
[8] Qi, Wenjing, Xue Liu, and Jing Zhao. Flower classification based on local and spatial
visual cues. Computer Science and Automation Engineering (CSAE), Vol. 3, 2012.
[9] Rassem, Taha H., and Bee Ee Khoo.Object class recognition using combination of
color SIFT descriptors.Imaging Systems and Techniques (IST), IEEE, 2011.
[10]Siraj, Fadzilah, Muhammad Ashraq Salahuddin, and Shahrul Azmi Mohd
Yusof.Digital Image Classification for Malaysian Blooming Flower. Computational
Intelligence, Modelling and Simulation (CIMSiM), IEEE, 2010.
[11]Saitoh, Takeshi, Kimiya Aoki, and Toyohisa Kaneko. Automatic recognition of
blooming flowers. Pattern Recognition, Vol. 1, 2004.
Kulkarni, T. and Uke, N. J., 2014. Implementation of Image based Flower
Classification System. International Journal of Computer Science and
Business Informatics, Vol. 13, No. 1, pp. 35-44.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 45
A Survey on Knowledge Analytics
of Text from Social Media
Dr. J. Akilandeswari
Professor and Head,
Department of Information Technology
Sona College of Technology,
Salem, India.
K. Rajalakshm
PG Scholar, Department of Information Technology
Sona College of Technology,
Salem, India.
ABSTRACT
Actionable knowledge discovery is a closed optimization problem solving process from problem
definition. It is used to extract the actionable data that are usable. Social media still contain many
comments that cannot be directly acted upon. If we could automatically filter out such noise and only
present actionable comments, decision making process will be easier. Automatically extracting
actionable knowledge from on line social media has been attracted a growing interest from both
academia and the industry. This paper gives a study in the systems and methods available text from
the social media like twitter or Facebook.
Keywords
knowledge discovery, social networking, classification.
1. INTRODUCTION
Social networking becomes one of the most important parts of our daily life.
It enables us to communicate with a lot of people. Social networking is
created to assist in online networking. These social sites are generally
communities created to support a common idea. Data mining is the process
of discovering actionable information from large sets of data. Actionable
knowledge discovery from user-generated content is a commodity much
sought after by industry and market research. The value of user-generated
content varies significantly from excellence to abuse. As the availability of
such content increases, identifying high-quality content in social sites based
on user contributions is very difficult. Social media sites become
increasingly important. In general social media demonstrate a rich variety of
information sources. In addition to the content itself, there is a large array of
non-content information obtainable in these sites, such as links between
items and unambiguous quality ratings from members of the community.
We argue that to achieve the goal we must gain a better understanding of
what actionable knowledge is, where it can be found and what kind of

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 46
language structures it contains. The aim of this work is to do so by
analyzing actionable knowledge in on-line social media conversation.
2. Related works
Maria Angela et al., [2] has proposed understanding Actionable knowledge
in social media BBC Question time and twitter. This paper will answer the
following questions: What is actionable knowledge, whether it can be
measured and where can we find for gaining better understanding of
actionable knowledge in twitter? There are three types of tweets: closed, re-
tweet, open. Actionable tweets can found in any of these categories. Three
steps are involved; 1) manually annotate the three subsets with action ability
scores. 2) Test the hypotheses by performing statistical annotated data. 3)
Use the W Matrix to automatically identify the language patterns in
actionable data. The method used in this paper prepares two sets Seta
containing actionable data and sets containing non actionable data. The two
sets of data are then loaded into the W matrix.
Eugene Agichtein et al., [3] have proposed to automatically asses the quality
of questions and answers provided by the user of the system. They take the
test case as Yahoo! Answers. They introduce the general classification
framework for combine the substantiation from different sources of
information, which can be adjusted automatically for a given social media
type and quality definition. Sub problem of quality evaluation is an essential
module for performing more advanced information retrieval tasks on the
question/answering. The interactions of users are organized around
questions like 1) asking a question 2) answering a question 3) selecting best
answers 4) voting on an answer.
Models:
• Intrinsic content quality: The content quality of each item. This is
mostly used text related.
• Punctuation and typos
• Syntactic and semantic complexity
• Grammatically
• Usage statistics: Clicks on the item.
Modeling content quality in community Question/Answering:
Application-specific user relationships:
The dataset, viewed as a graph, contains multiple types of nodes and
multiple types of interactions

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 47
Fig 2.1 Partial Entity-Relationship Diagram for Answers
The relationships between questions, user asking and answering
questions, and answers can be captured by a tripartite graph outlined in the
figure where an edge represents an explicit relationship between the
different node types. Since a user is not allowed to answer his/her own
questions.
Fig 2.2 Interaction of user-questions-answers modeled as a Tri-partiate Graph.
The types of features on the question sub tree:
Q represents features from the question being answered.
QU represents features from the asker of the question being
answered.
QA represents features from the other answer to the same question.

IJCSBI.ORG
ISSN: 1694-2108 | Vol. 13, No. 1. MAY 2014 48
Fig 2.3 Types of features available for inferring the quality of question.
The types of features on the user sub tree:
UA represents features from the answers of the user
UQ represents features from the question of the user
UV represents features from the votes of the user
UQA represents features from answers user received to the user’s
question.
U represents other user based features.
Fig 2. 4 Types features available for inferring the quality of a question
A represents feature directly from the answer received.
AU represents features from the answers from the question being answered.

Vol 13 No 1 - May 2014

Recommended

Recommended

More Related Content

What's hot

What's hot (18)

Similar to Vol 13 No 1 - May 2014

Similar to Vol 13 No 1 - May 2014 (20)

More from ijcsbi

More from ijcsbi (20)

Recently uploaded

Recently uploaded (20)

Vol 13 No 1 - May 2014